uql-orm 0.18.0 → 0.20.0

package/dist/dialect/abstractSqlDialect.js

@@ -1,10 +1,12 @@
 import { getMeta } from '../entity/index.js';
-import { resolveVectorCast } from '../schema/canonicalType.js';
 import { QueryRaw, RAW_ALIAS, RAW_VALUE, } from '../type/index.js';
-import { applyFilters, buildQueryWhereAsMap, buildSortMap, escapeSqlId, fillOnFields, filterFieldKeys, flatObject, getFieldKeys, getInsertFieldKeys, getKeys, getRelationRequestSummary, getSoftDeleteValue, hasKeys, hasMultipleKeys, isJsonType, isPopulatingRelations, isVectorSearch, normalizeScalarFieldSelection, parseGroupMap, parseRelationAtKey, raw, withoutSoftDeleteFilter, } from '../util/index.js';
-import { AbstractDialect } from './abstractDialect.js';
+import { escapeSingleQuotes } from '../util/ansiSqlLiteral.js';
+import { applyFilters, asSelectMap, buildQueryWhereAsMap, buildSortMap, escapeSqlId, fillOnFields, filterFieldKeys, flatObject, getFieldKeys, getInsertFieldKeys, getKeys, getRelationRequestSummary, getSoftDeleteValue, hasKeys, hasMultipleKeys, isJsonType, isJsonUpdateOp, isOperatorObject, isOperatorOnlyObject, isPopulatingRelations, isVectorSearch, normalizeScalarFieldSelection, parseGroupMap, parseRelationAtKey, raw, someValue, withoutSoftDeleteFilter, } from '../util/index.js';
+import { buildElemMatchConditions } from './jsonArrayElemMatchUtils.js';
+import { jsonElemExists } from './jsonSql.js';
 import { SqlQueryContext } from './queryContext.js';
-export class AbstractSqlDialect extends AbstractDialect {
+import { VectorSqlDialect } from './vectorSqlDialect.js';
+export class AbstractSqlDialect extends VectorSqlDialect {
     isolationLevelStrategy = 'inline';
     alterColumnStrategy = 'single-statement';
     alterColumnSyntax = 'ALTER COLUMN';
@@ -17,12 +19,6 @@ export class AbstractSqlDialect extends AbstractDialect {
      * `insertMany` splits larger batches into multiple statements based on this limit.
      */
     maxBindValues = 32766;
-    /** Vector index operator classes, keyed by distance metric. Partial: not every dialect supports every metric. */
-    vectorOpsClass = undefined;
-    vectorExtension = undefined;
-    get escapeIdChar() {
-        return this.quoteChar;
-    }
     getBeginTransactionStatements(isolationLevel) {
         const level = isolationLevel?.toUpperCase();
         const strategy = this.isolationLevelStrategy;
@@ -78,7 +74,7 @@ export class AbstractSqlDialect extends AbstractDialect {
     search(ctx, entity, q = {}, opts = {}) {
         const meta = getMeta(entity);
         const tableName = this.resolveTableName(entity, meta);
-        const prefix = this.resolveRelationAwarePrefix(tableName, meta, opts, q.$select, q.$populate);
+        const prefix = this.resolveRelationAwarePrefix(tableName, meta, opts, asSelectMap(q.$select), q.$populate);
         if (opts.prefix !== prefix) {
             opts = { ...opts, prefix };
         }
@@ -93,11 +89,11 @@ export class AbstractSqlDialect extends AbstractDialect {
         let selectArr;
         if (select) {
             if (Array.isArray(select)) {
-                // Internal-only path: raw SQL expressions passed as QueryRaw[]
+                // Raw SQL projections passed as QueryRaw[]
                 selectArr = select;
             }
             else {
-                selectArr = normalizeScalarFieldSelection(meta, select, exclude);
+                selectArr = normalizeScalarFieldSelection(meta, asSelectMap(select), exclude);
             }
             const id = meta.id;
             if (id && opts.prefix && !selectArr.includes(id)) {
@@ -148,7 +144,7 @@ export class AbstractSqlDialect extends AbstractDialect {
     select(ctx, entity, select, exclude, populate, opts = {}, distinct, sort) {
         const meta = getMeta(entity);
         const tableName = this.resolveTableName(entity, meta);
-        const mapSelect = Array.isArray(select) ? undefined : select;
+        const mapSelect = asSelectMap(select);
         const prefix = this.resolveRelationAwarePrefix(tableName, meta, opts, mapSelect, populate);
         ctx.append(distinct ? 'SELECT DISTINCT ' : 'SELECT ');
         this.selectFields(ctx, entity, select, { prefix }, exclude);
@@ -392,6 +388,17 @@ export class AbstractSqlDialect extends AbstractDialect {
         ['$like', (v) => v],
         ['$ilike', (v) => v.toLowerCase()],
     ]);
+    /**
+     * The case-insensitive `LIKE_OP_MAP` keys - the value is lowercased, so the comparison must use
+     * `ilikeExpr` (Postgres's `ILIKE`) rather than `LIKE`. `$includes` is deliberately excluded even
+     * though it starts with the substring `$i`: it is case-sensitive, unlike `$iincludes`.
+     */
+    static LIKE_CASE_INSENSITIVE_OPS = new Set([
+        '$istartsWith',
+        '$iendsWith',
+        '$iincludes',
+        '$ilike',
+    ]);
     /** Builds `prefix.column` from an already-resolved field. */
     columnWithPrefix(key, field, prefix) {
         const columnName = this.resolveColumnName(key, field);
@@ -470,7 +477,7 @@ export class AbstractSqlDialect extends AbstractDialect {
         }
     }
     appendLikeOp(ctx, field, op, wrappedVal) {
-        const isIlike = op.startsWith('$i') || op === '$ilike';
+        const isIlike = AbstractSqlDialect.LIKE_CASE_INSENSITIVE_OPS.has(op);
         const ph = this.addValue(ctx.values, wrappedVal);
         if (isIlike && field) {
             ctx.append(this.ilikeExpr(field, ph));
@@ -504,15 +511,17 @@ export class AbstractSqlDialect extends AbstractDialect {
      * Used by both `$elemMatch` and dot-notation paths.
      * All dialect-specific behavior comes from overridable methods on `this`.
      */
-    buildJsonFieldCondition(ctx, fieldAccessor, jsonPath, op, value) {
+    buildJsonFieldCondition(ctx, fieldAccessor, jsonPath, op, value, asJson = false) {
         const jsonField = fieldAccessor(jsonPath);
         switch (op) {
             case '$eq':
-                return value === null ? `${jsonField} IS NULL` : `${jsonField} = ${this.addValue(ctx.values, value)}`;
+                if (value === null)
+                    return `${jsonField} IS NULL`;
+                return `${this.jsonComparand(jsonField, value)} = ${this.jsonOperand(ctx, value, asJson)}`;
             case '$ne':
                 if (value === null)
                     return `${jsonField} IS NOT NULL`;
-                return this.neExpr(jsonField, this.addValue(ctx.values, value));
+                return this.neExpr(this.jsonComparand(jsonField, value), this.jsonOperand(ctx, value, asJson));
             case '$gt':
                 return `${this.numericCast(jsonField)} > ${this.addValue(ctx.values, value)}`;
             case '$gte':
@@ -541,7 +550,7 @@ export class AbstractSqlDialect extends AbstractDialect {
                 return `${jsonField} ${this.regexpOp} ${this.addValue(ctx.values, value)}`;
             case '$in':
             case '$nin':
-                return this.jsonInNin(ctx, jsonField, op, value);
+                return this.jsonInNin(ctx, jsonField, op, value, asJson);
             case '$all':
                 return this.jsonAll(ctx, jsonField, value);
             case '$size':
@@ -552,20 +561,117 @@ export class AbstractSqlDialect extends AbstractDialect {
                 throw TypeError(`unknown operator: ${op}`);
         }
     }
-    jsonInNin(ctx, jsonField, op, value) {
-        return `${jsonField}${this.formatIn(ctx, Array.isArray(value) ? value : [], op === '$nin')}`;
+    jsonInNin(ctx, jsonField, op, value, asJson) {
+        const values = Array.isArray(value) ? value : [];
+        const negate = op === '$nin';
+        if (!asJson) {
+            return `${this.jsonComparand(jsonField, values)}${this.formatIn(ctx, values, negate)}`;
+        }
+        // JSON values have no portable array literal, so the set expands into explicit comparisons.
+        const comparisons = values.map((val) => `${jsonField} ${negate ? '<>' : '='} ${this.jsonScalarParam(ctx, val)}`);
+        return `(${comparisons.join(negate ? ' AND ' : ' OR ')})`;
+    }
+    /** The bound operand of a JSON comparison: JSON-encoded when comparing against the JSON value. */
+    jsonOperand(ctx, value, asJson) {
+        return asJson ? this.jsonScalarParam(ctx, value) : this.addValue(ctx.values, value);
+    }
+    /**
+     * The left side of a comparison against a JSON scalar, cast when the operand is numeric - see
+     * {@link jsonCompareMode} for why each mode exists.
+     */
+    jsonComparand(jsonField, value) {
+        return this.jsonCompareMode(value) === 'numeric' ? this.numericCast(jsonField) : jsonField;
     }
-    jsonAll(ctx, jsonField, value) {
-        throw TypeError(`$all is not supported in the base SQL dialect - override in dialect subclass`);
+    /**
+     * How a JSON scalar has to be compared against `value` (or, for `$in`/`$nin`, against every element
+     * of it). Extracting a JSON value yields *text*, which loses the type, so each operand type is
+     * compared in the representation every engine agrees on:
+     * - `numeric` - cast the accessor. Keeps `1` equal to a stored `1.0`, which strict JSON equality
+     *   would not, and satisfies drivers that send typed parameters (`text = integer` otherwise).
+     * - `json` - compare the JSON value against a JSON-encoded parameter. No cast recovers a boolean
+     *   portably: PostgreSQL raises `text = boolean` and MySQL matches `'true'` against `1`.
+     * - `text` - compare as extracted, which is also what the string operators need.
+     *
+     * Mixed operand types fall back to `text`, since one comparison cannot be two shapes at once.
+     */
+    jsonCompareMode(value) {
+        const operands = Array.isArray(value) ? value : [value];
+        if (operands.length === 0) {
+            return 'text';
+        }
+        if (operands.every((operand) => typeof operand === 'boolean')) {
+            return 'json';
+        }
+        return operands.every((operand) => typeof operand === 'number') ? 'numeric' : 'text';
     }
-    jsonSize(ctx, jsonField, value) {
-        throw TypeError(`$size is not supported in the base SQL dialect - override in dialect subclass`);
+    /**
+     * Whether the dialect's array containment ({@link jsonAll}) matches an object element that merely
+     * *includes* the given keys, as PostgreSQL's `@>` and MySQL's `JSON_CONTAINS` do. SQLite compares
+     * elements as whole JSON text, so it cannot express a partial match and always expands the
+     * per-field form below.
+     */
+    jsonContainmentIsPartial = true;
+    /**
+     * Whether an exploded *scalar* element keeps its SQL type. SQLite's `json_each` yields JSON
+     * booleans as `0`/`1` integers and numbers as numbers, so such an element compares directly to a
+     * bound value; PostgreSQL and MySQL explode scalars to text, losing the type, so a non-string
+     * operand there has to compare as JSON (see {@link isJsonbOp}).
+     */
+    jsonScalarElemKeepsType = false;
+    /**
+     * `$elemMatch`: at least one element of the JSON array satisfies `match`. Three shapes, decided
+     * here so every dialect only supplies {@link jsonElemFrom} / {@link jsonElemRef}:
+     * - keys are operators (`{ $startsWith: 'ad' }`) - scalar elements, conditions on the element;
+     * - a plain object with no nested operators - containment, which is the only form an index serves;
+     * - otherwise - per-field conditions over the exploded objects.
+     */
+    jsonElemMatch(ctx, jsonField, match) {
+        // Conditions on the element itself. One `FROM` serves them all, so the element is read as JSON
+        // only when *every* operand needs it - the same all-operands rule the comparison classifier uses.
+        if (isOperatorOnlyObject(match)) {
+            const entries = Object.entries(match);
+            const asJson = !this.jsonScalarElemKeepsType && entries.every(([op, val]) => this.isJsonbOp(op, val));
+            const conditions = entries.map(([op, val]) => this.buildJsonFieldCondition(ctx, () => this.jsonElemRef(undefined, asJson), '', op, val, asJson));
+            return jsonElemExists(this.jsonElemFrom(jsonField, [], asJson), conditions);
+        }
+        if (isOperatorObject(match)) {
+            throw TypeError(`$elemMatch cannot mix operators with field names: ${Object.keys(match).join(', ')}`);
+        }
+        // A plain object with no nested operators is containment, which is also the only form an index
+        // can serve. SQLite compares elements exactly, so it always expands the per-field form below.
+        if (this.jsonContainmentIsPartial && !someValue(match, isOperatorObject)) {
+            return this.jsonAll(ctx, jsonField, [match]);
+        }
+        const conditions = buildElemMatchConditions(match, (field, op, opVal) => {
+            const asJson = this.isJsonbOp(op, opVal);
+            return this.buildJsonFieldCondition(ctx, (f) => this.jsonElemRef(f, asJson), field, op, opVal, asJson);
+        });
+        return jsonElemExists(this.jsonElemFrom(jsonField, Object.keys(match)), conditions);
     }
-    jsonElemMatch(ctx, jsonField, value) {
-        throw TypeError(`$elemMatch is not supported in the base SQL dialect - override in dialect subclass`);
+    /**
+     * Whether the operator reads the JSON *value* instead of its text form. The array operators always
+     * do. Equality joins them for boolean operands, because extracting JSON as text loses the type in
+     * a way no cast recovers portably: PostgreSQL raises `operator does not exist: text = boolean`,
+     * MySQL compares `'true'` to `1` and silently matches nothing, and SQLite's `json_extract` yields
+     * `1`. Comparing the JSON value against a JSON-encoded parameter is exact on every dialect.
+     *
+     * Numbers stay on the text accessor with a numeric cast ({@link jsonComparand}), which keeps
+     * `1` equal to `1.0` - JSON equality would not.
+     */
+    isJsonbOp(op, value) {
+        if (op === '$all' || op === '$size' || op === '$elemMatch') {
+            return true;
+        }
+        const comparesValue = op === '$eq' || op === '$ne' || op === '$in' || op === '$nin';
+        return comparesValue && this.jsonCompareMode(value) === 'json';
     }
-    isJsonbOp(op) {
-        return op === '$all' || op === '$size' || op === '$elemMatch';
+    /** A JSON-encoded bound parameter, cast to the dialect's JSON type. */
+    jsonScalarParam(ctx, value) {
+        if (value instanceof QueryRaw) {
+            return this.addValue(ctx.values, value);
+        }
+        ctx.pushValue(JSON.stringify(value));
+        return this.getJsonCastExpr();
     }
     getComparisonKey(ctx, entity, key, { prefix } = {}) {
         const meta = getMeta(entity);
@@ -632,59 +738,6 @@ export class AbstractSqlDialect extends AbstractDialect {
             ctx.append(this.escapeId(name) + direction);
         });
     }
-    /**
-     * Resolve common parameters for a vector similarity ORDER BY expression.
-     * Shared by all dialect overrides of `appendVectorSort`.
-     */
-    resolveVectorSortParams(meta, key, search) {
-        const field = meta.fields[key];
-        const colName = this.resolveColumnName(key, field);
-        const distance = search.$distance ?? field?.distance ?? 'cosine';
-        const vectorCast = resolveVectorCast(field);
-        return { colName, distance, field, vectorCast };
-    }
-    /**
-     * Mapping of UQL vector distance metrics to native SQL functions.
-     * Override in dialects that use function-call syntax (e.g. SQLite, MariaDB).
-     * Dialects with operator-based syntax (e.g. Postgres) leave this empty and override `appendVectorSort` directly.
-     */
-    vectorDistanceFns = new Map();
-    /**
-     * Append a vector similarity function call: `fn(col, ?)`.
-     * Used by dialects that express vector distance via SQL functions (SQLite, MariaDB).
-     */
-    appendFunctionVectorSort(ctx, meta, key, search, dialectName) {
-        const { colName, distance, vectorCast } = this.resolveVectorSortParams(meta, key, search);
-        const fn = this.vectorDistanceFns.get(distance);
-        if (!fn) {
-            throw Error(`${dialectName} does not support vector distance metric: ${distance}`);
-        }
-        ctx.append(`${fn}(${this.escapeId(colName)}, `);
-        ctx.addValue(`[${search.$vector.join(',')}]`);
-        if (vectorCast && dialectName === 'PostgreSQL') {
-            ctx.append(`::${vectorCast}`);
-        }
-        ctx.append(')');
-    }
-    /**
-     * Append a vector distance projection.
-     */
-    appendVectorProjection(ctx, meta, key, search) {
-        this.appendVectorSort(ctx, meta, key, search);
-        ctx.append(` AS ${this.escapeId(search.$project)}`);
-    }
-    /**
-     * Append a vector similarity ORDER BY expression.
-     * Default: auto-delegates to `appendFunctionVectorSort` when `vectorDistanceFns` has entries.
-     * Override for operator-based syntax (e.g. PostgreSQL `<=>`, `<->` operators).
-     */
-    appendVectorSort(ctx, meta, key, search) {
-        if (this.vectorDistanceFns.size > 0) {
-            this.appendFunctionVectorSort(ctx, meta, key, search, this.dialectName);
-            return;
-        }
-        throw new TypeError('Vector similarity sort is not supported by this dialect. Use raw() for vector queries.');
-    }
     pager(ctx, opts) {
         if (opts.$limit) {
             ctx.append(` LIMIT ${Number(opts.$limit)}`);
@@ -894,8 +947,8 @@ export class AbstractSqlDialect extends AbstractDialect {
             const columnName = this.resolveColumnName(key, field);
             const escapedCol = this.escapeId(columnName);
             const value = filledPayload[key];
-            if (this.isJsonUpdateOp(value)) {
-                this.formatJsonUpdate(ctx, escapedCol, value);
+            if (isJsonUpdateOp(value)) {
+                this.formatJsonUpdate(ctx, escapedCol, value, field);
             }
             else {
                 ctx.append(`${escapedCol} = `);
@@ -1019,50 +1072,59 @@ export class AbstractSqlDialect extends AbstractDialect {
         ctx.addValue(value);
     }
     /**
-     * Generate SQL for a JSONB merge and/or unset operation.
-     * Called from `update()` when a field value has `$merge`, `$unset`, and/or `$push` operators.
-     * Generates the full `"col" = <expression>` assignment.
-     *
-     * Base implementation uses MySQL-compatible syntax with *shallow* merge semantics
-     * (RHS top-level keys replace LHS top-level keys, matching PostgreSQL's `jsonb || jsonb`).
-     * Override in dialect subclasses when a dialect needs different JSON function semantics.
+     * Reads `operand` as a JSON value. Passing the `?` placeholder yields the cast for a bound
+     * parameter, and passing an expression re-reads text as JSON - the same SQL either way, which is
+     * why this is one hook rather than a placeholder variant plus an expression variant.
+     */
+    jsonCast(operand) {
+        return `CAST(${operand} AS JSON)`;
+    }
+    /**
+     * The JSON cast applied to a bound placeholder. Only the positional-placeholder dialects use it -
+     * PostgreSQL binds JSON through {@link PgLikeSqlDialect.jsonScalarParam} instead.
      */
     getJsonCastExpr() {
-        return 'CAST(? AS JSON)';
+        return this.jsonCast('?');
     }
-    formatJsonUpdate(ctx, escapedCol, value) {
+    /**
+     * Generate the full `"col" = <expression>` assignment for a JSON update operator payload.
+     * Called from `update()` when a field value is a {@link JsonUpdateOp}.
+     *
+     * Each operator wraps the expression built so far, innermost-first in the order stated on
+     * {@link JsonUpdateOp} (`$pull` -> `$set` -> `$push` -> `$unset`), so dialects only supply the
+     * four SQL fragments below. Two invariants keep every dialect consistent and keep bound values in
+     * step with their placeholders:
+     * - `$pull` is innermost and its subquery reads `escapedCol`, so its value binds exactly once.
+     * - Later fragments reference `expr` at most once, so a `$pull` subquery is never duplicated
+     *   (which would bind its value twice on positional-placeholder dialects). PostgreSQL's `$push`
+     *   is the one exception, and is safe there because its placeholders are numbered.
+     */
+    formatJsonUpdate(ctx, escapedCol, value, field) {
+        // Centralizes the one narrowing cast: the payload's keys are typed against the entity's JSON
+        // payload, which the dialects do not need - they only build SQL from keys and values.
+        const { $pull, $set, $push, $unset } = value;
         let expr = escapedCol;
-        if (hasKeys(value.$merge)) {
-            const merge = value.$merge;
-            expr = `JSON_SET(COALESCE(${escapedCol}, '{}')`;
-            for (const [key, v] of Object.entries(merge)) {
-                expr += `, '$.${this.escapeJsonKey(key)}', ${this.getJsonCastExpr()}`;
-                ctx.pushValue(JSON.stringify(v));
-            }
-            expr += ')';
-        }
-        if (hasKeys(value.$push)) {
-            const push = value.$push;
-            expr = `JSON_ARRAY_APPEND(${expr}`;
-            for (const [key, v] of Object.entries(push)) {
-                expr += `, '$.${this.escapeJsonKey(key)}', ${this.getJsonCastExpr()}`;
-                ctx.pushValue(JSON.stringify(v));
-            }
-            expr += ')';
+        if (hasKeys($pull)) {
+            expr = this.jsonPull(ctx, expr, escapedCol, $pull);
         }
-        if (value.$unset?.length) {
-            for (const key of value.$unset) {
-                expr = `JSON_REMOVE(${expr}, '$.${this.escapeJsonKey(key)}')`;
-            }
+        if (hasKeys($set)) {
+            expr = this.jsonSet(ctx, expr, $set, field);
+        }
+        if (hasKeys($push)) {
+            expr = this.jsonPush(ctx, expr, $push);
+        }
+        if ($unset?.length) {
+            expr = this.jsonUnset(ctx, expr, $unset);
         }
         ctx.append(`${escapedCol} = ${expr}`);
     }
-    isJsonUpdateOp(value) {
-        return typeof value === 'object' && value !== null && ('$merge' in value || '$unset' in value || '$push' in value);
-    }
-    /** Escapes a JSON key for safe interpolation into SQL string literals. */
-    escapeJsonKey(key) {
-        return key.replace(/'/g, "''");
+    /**
+     * Remove every element equal to the given value, per array key. Each key wraps the expression
+     * built so far, so dialects only supply {@link jsonPullKey} - and because every key reads
+     * `escapedCol` rather than the accumulated expression, values bind once, in key order.
+     */
+    jsonPull(ctx, expr, escapedCol, pull) {
+        return Object.entries(pull).reduce((acc, [key, value]) => this.jsonPullKey(ctx, acc, escapedCol, key, value), expr);
     }
     getRawValue(ctx, opts) {
         const { value, prefix = '', escapedPrefix, autoPrefixAlias } = opts;
@@ -1127,7 +1189,8 @@ export class AbstractSqlDialect extends AbstractDialect {
         operators.forEach((op, index) => {
             if (index > 0)
                 ctx.append(' AND ');
-            const sql = this.buildJsonFieldCondition(ctx, (f) => accessor(this.isJsonbOp(op)), jsonPath, op, value[op]);
+            const asJson = this.isJsonbOp(op, value[op]);
+            const sql = this.buildJsonFieldCondition(ctx, () => accessor(asJson), jsonPath, op, value[op], asJson);
             if (sql) {
                 ctx.append(sql);
             }
@@ -1146,7 +1209,7 @@ export class AbstractSqlDialect extends AbstractDialect {
         let expr = escapedColumn;
         for (let i = 0; i < segments.length; i++) {
             const op = i === segments.length - 1 ? '->>' : '->';
-            expr = `(${expr}${op}'${this.escapeJsonKey(segments[i])}')`;
+            expr = `(${expr}${op}'${escapeSingleQuotes(segments[i])}')`;
         }
         return expr;
     }
@@ -1154,7 +1217,7 @@ export class AbstractSqlDialect extends AbstractDialect {
         const segments = jsonPath.split('.');
         let expr = escapedColumn;
         for (const segment of segments) {
-            expr = `(${expr}->'${this.escapeJsonKey(segment)}')`;
+            expr = `(${expr}->'${escapeSingleQuotes(segment)}')`;
         }
         return expr;
     }